hesela.devAssise
Assise is a research distributed file system that combines client-local persistent-memory caches with replicated coherence and linearizable file I/O.
LibFS intercepts supported calls under a configured namespace. SharedFS coordinates leases; running with one local node is a documented configuration, not replicated protection. [2]
In the paper's pessimistic mode, fsync waits for replication. Optimistic mode may defer replication until dsync or eviction, risking data loss when local recovery is impossible. Both preserve a recoverable prefix, not necessarily every completed write after node loss. [1]
Each replica flushes received writes with CLWB and SFENCE before acknowledging replication. RDMA transfer completion alone is not this persistence guarantee. [1]
The pinned mlfs_posix_fsync wrapper calls mlfs_do_rsync only when DISTRIBUTED is compiled in, then returns zero. This wrapper alone does not establish the full recovery contract. [3]
Relations
- related
- Persistent memory, Chain replication, RDMA storage networking, Optimistic crash consistency, Crash consistency
Citations
- Anderson et al. Assise: Performance and Availability via Client-local NVM in a Distributed File System. OSDI 2020, pp. 1011-1027, sections 3-5.
- Assise authors. README at c191f20b26a86bef79bb8c943b33c0c75e17c802: namespace interception, cluster configuration and build flags.
- Assise authors. posix_interface.c at c191f20b26a86bef79bb8c943b33c0c75e17c802, mlfs_posix_fsync.
Further reading
Sources & caveats
- Source synthesis of the OSDI 2020 design and a pinned public artifact, not an executed prototype, hardware measurement or compatibility recommendation.
- Paper guarantees and an inspected wrapper are different evidence levels; a return value is not a fault-injection result.